A Graph Convolutional Network-Based Approach for Dynamic Connectivity Prediction in 5G Networks

A Graph Convolutional Network-Based Approach for Dynamic Connectivity Prediction in 5G Networks Raul Barbosa∗§, Marco Ara´ ujo∗, Petia Georgieva†‡, Susana Sargento†§, Pedro Rito†§ ∗Capgemini Engineering, Porto, Portugal †Departamento de Electr´ onica, Telecomunicac¸ ˜ oes e Inform´ atica, Universidade de Aveiro, Portugal ‡Institute of Electronics and Informatics Engineering of Aveiro (IEETA) §Instituto de Telecomunicac¸ ˜ oes, Universidade de Aveiro, Campus Universit´ ario de Santiago, 3810-193, Aveiro, Portugal Abstract—Next-generation beyond 5G networks face significant challenges in ensuring resilient connectivity and lowlatency performance for critical applications in dynamic and dense environments. This paper presents a Graph Convolutional Network (GCN)-based model to optimize handover decisions by predicting the most suitable gNodeB (gNB) for user connection in real-time. Leveraging historical connectivity data and network conditions, the model forecasts gNB connectivity and incorporates a threshold mechanism to reduce unnecessary handovers and mitigate the ”ping-pong” effect. A graph representation of a real 5G dataset is constructed, where nodes represent gNBs with connectivity attributes, and edges capture potential handovers weighted by connectivity differences. The results demonstrate that the proposed GCN model improves the network resilience by ensuring stable connectivity and minimizing disruptions, achieving enhanced user experience without increasing handover frequency. This study underscores the potential of machine learning-driven resilience mechanisms in next-generation networks, offering a robust framework for dynamic connectivity management in high-mobility and latency-sensitive scenarios. Index Terms—5G, Next-Generation Networks, Resilience, Graph Convolutional Networks, gNB Handovers I. INTRODUCTION The introduction of 5G telecommunication networks has led to significant advancements in network connectivity and performance management, emphasizing data resilience and reduced latency. These developments bring both challenges and opportunities in optimizing network resources, particularly for applications that demand high performance and reliability, such as real-time communication and mission-critical IoT services. Traditional handover mechanisms, which rely on metrics like signal strength or power levels, are insufficient in modern networks where factors such as latency, jitter, and bandwidth variability play critical roles. To address these limitations, more adaptive and intelligent approaches are required. In this context, Artificial Intelligence (AI), specifically through Graph Neural Networks (GNNs), offers a promising solution to optimize handovers between 5G gNodeBs. GNNs This work was supported by the European Union’s Horizon Europe research and innovation program through the project IMAGINE-B5G under grant agreement No. 101096452; use graph-based data structures to model the complex relationships and dependencies between network nodes, which traditional handover approaches may overlook [1]. By representing the network as a graph, where nodes are gNBs and edges denote the potential handover relationships, GNNs can leverage the connectivity and interaction between nodes to inform handover decisions. Graph Convolutional Networks (GCNs), a subclass of GNNs, enable advanced modeling by applying convolutional operations to graph data. GCNs aggregate information from neighboring nodes, making them particularly effective for predicting connectivity and optimizing handover decisions in dynamic network environments. In modern 5G networks, graph-based methods have shown potential in improving tasks such as feature extraction for deep reinforcement learning (DRL) agents, routing optimization, resource allocation, and virtual network function (VNF) deployment [2]. GNN-based approaches capture the growing complexity of user-centric ultra-dense networks and IoT deployments, recognizing the significance of multi-scale relationships between network nodes [3], [4]. This paper presents a GCN-based approach to enhance resilience in 5G networks by predicting connectivity for the next time step (t+ 1) and improving handover decisions. The findings demonstrate that this model reduces unnecessary handovers, enhances connectivity stability, and supports latencysensitive applications. The remainder of this paper is organized as follows. Section II provides an overview of related work in the field. Section III describes the methodology and the proposed GCN-based model. Section IV presents the experimental setup and results, followed by a discussion of the findings. Finally, Section V concludes the paper and outlines directions for future research. II. RELATED WORK The domain of network management and optimization has undergone a significant transformation with the advent of AI and machine learning techniques. Among these, GNNs have garnered substantial interest for their ability to model complex relationships within network structures. This section reviews the relevant literature, providing a comprehensive overview of GNN architectures and their applications, the unique challenges and requirements of 5G networks, and the role of GNNs in optimizing the network performance. A. GNN and GCN Architectures GNNs are designed to process graph-structured data, making them particularly effective for applications where relationships between entities are critical [4]. They have been successfully applied to tasks such as node classification, link prediction, and graph classification, capturing dependencies between nodes and edges [2], [3]. A key subclass, Graph Convolutional Networks (GCNs), extends Convolutional Neural Networks (CNNs) to graphs by aggregating and transforming information from neighboring nodes. This enables GCNs to effectively learn graph topologies and features, excelling in tasks like node classification and link prediction. The foundational work in [5] demonstrated their efficacy in semi-supervised learning on graph-structured data. B. 5G Networks and the Role of Handovers The introduction of 5G networks has revolutionized telecommunications by providing higher data rates, lower latency, and enhanced connectivity. These capabilities are critical for applications like autonomous driving and real-time video streaming, but also introduce challenges in resource management and maintaining seamless connectivity. Handovers, where mobile devices transition between gNBs, are a critical aspect of 5G network management. Their complexity has increased due to the higher density of gNBs and the diverse network infrastructure. Traditional mechanisms based on static thresholds like signal strength, are insufficient for addressing 5G’s dynamic requirements, as they often overlook latency and bandwidth considerations essential for modern applications [6]. To address these challenges, handover management in 5G must integrate connectivity metrics and performance factors, such as latency, jitter and bandwidth. A more nuanced approach involves not only allocating resources, but also optimizing network parameters to meet quality of service (QoS) requirements, ensuring reliable performance for diverse applications. C. GCNs in the Context of Communication Infrastructure GCNs offer a promising solution to the challenges of handover management in 5G networks. By modeling the network as a graph, where nodes represent gNBs and edges represent potential transitions, GCNs can learn from historical connectivity data and current network conditions to predict the optimal gNB for handover. This approach leverages the ability of GCNs to capture the complex dependencies and interactions within the network, enabling more accurate and dynamic decision-making [2]. Recent studies highlight GCNs’ potential in network management. Tam et al. [2] reviewed the use of GNNs in tasks such as routing, resource allocation, and service orchestration in complex network architectures. Their findings emphasize that GNNs effectively handle the dynamic, non-Euclidean nature of network topologies, enabling more efficient and autonomous network operations. GCNs have been applied to Multi-access Edge Computing (MEC). Leng et al. [7] introduced a GCN-based reinforcement learning method for task offloading, which significantly reduced server idle time and improved task scheduling efficiency, demonstrating the GCNs’ ability to optimize resource allocation in dynamic environments. Similarly, Zhang et al. [8] developed a graph-based load balancing method for 5G/6G networks, achieving high service acceptance rates while maintaining network stability. Their approach showcases the scalability and effectiveness of GCNs in managing large-scale, dynamic network scenarios. Chen et al. [3] proposed an actor-critic mechanism leveraging GCNs to optimize multitask offloading in edge computing. Their results demonstrated substantial reductions in latency and energy consumption, validating the practical value of GCNs in performance-critical scenarios. Building on these advancements, our study introduces a novel application of GCNs specifically tailored for dynamic gNB connectivity forecasting in 5G networks. While previous research has shown the power of GCNs in general network management tasks, our approach focuses on leveraging historical connectivity data and current network conditions to predict the optimal gNB for handover. This targeted application addresses the critical challenge of handover management in 5G environments, where real-time decision-making and accurate predictions are crucial. By integrating GCNs into the handover process, our methodology provides a more efficient and responsive solution, building on prior studies to offer a practical contribution to next-generation telecommunications. III. METHODOLOGY Our research focuses on GCNs to optimize network connectivity in 5G environments by predicting the optimal gNB for the user connection. Our approach forecasts connectivity at the next time step (t+ 1) using historical data and current network conditions, enabling informed and dynamic handover decisions. This is particularly critical to enhance the network performance in high-mobility and latency-sensitive scenarios. The proposed method models the network as a graph, where nodes represent gNBs and edges denote potential transitions weighted by connectivity metrics. The GCN-based model is trained to predict the future connectivity and select the optimal gNB, directly impacting the network efficiency and user experience. To validate this approach, we utilize a public 5G dataset containing detailed gNB connectivity attributes. This dataset is used to construct the graph and train the model, capturing relationships and dependencies between gNBs to enable anticipatory handover decisions. A. Data Collection and Analysis We utilized a public 5G dataset collected through the Aveiro Tech City Living Lab (ATCLL) infrastructure in Aveiro and a commercial cellular network [9], [10]. The dataset includes 17 days of connectivity data from a public bus equipped with an On-Board Unit (OBU). Table I summarizes the key features used for this study. TABLE I DESCRIPTION OF FEATURES Feature Description RSRP, RSRQ, SNR Signal strength and quality metrics RTT Round-Trip Time (ms) Packet Loss Percentage of lost packets (%) Bitrate Transmission speed (Mbps) Jitter Variation in packet delay (ms) Timestamp Time of data capture To address outliers and irregularities, we removed values outside the Interquartile Range (IQR) and zero RTT values. Weekend samples were discarded due to limited data, and the dataset was resampled at 60-second intervals. Missing values were handled using interpolation and reference data from corresponding time slots in other weeks. We normalized features such as RSRP, Bitrate, Packet loss, and RTT to a [0,1] range using the MinMaxScaler. To quantify the network quality, we calculate a new feature, referred to as the ”connectivity,” which combines multiple connectivity metrics into a single value. This score is computed using a weighted sum of normalized values for RSRP, bitrate, RTT, and packet loss. The general formula used to calculate the connectivity is: connectivity =wrsrp ·rsrp +wbitrate ·bitrate +wRTT ·(1 −RTT) + wpacket loss ·(1 −packet loss)(1) In this formula, wrsrp,wbitrate,wRTT, and wpacket loss represent the weights assigned to each metric. Initially, we adopted a weight configuration where RTT was assigned a higher weight of 0.4, while RSRP, Bitrate, and Packet loss were each assigned weights of 0.2. This configuration reflects the significant impact of RTT on latency-sensitive applications, with (1 −RTT)and (1 −packet loss)emphasizing that lower RTT and Packet loss values are preferable, given their inverse relationship with network quality. The selection of these weights was based on an empirical study, discussed in Section IV, where multiple configurations were tested to evaluate their impact on connectivity stability, latency and handover frequency. By emphasizing RTT with a higher weight, the connectivity value becomes more sensitive to latency variations, an essential factor for real-time applications in 5G environments. This connectivity is then calculated for each timestamp, with additional columns created to store the current connectivity (connectivity(t)) and the score from the previous timestamp (connectivity(t-1)). This temporal information allows the model to account for recent connectivity trends when making handover decisions, ensuring the selection of the optimal gNB for the next time step. Figure 1 illustrates the behavior of the connectivity over the entire dataset. Fig. 1. Connectivity behavior across the dataset The mean connectivity is approximately 0.61, indicating a moderate level of connectivity across the dataset. The standard deviation of 0.20 reveals a fair amount of variability in connectivity, which may reflect the changing network conditions or mobility of the user. The minimum recorded connectivity is 0.05, while the maximum reaches 0.92, highlighting the range of connectivity experiences. This wide range underscores the network’s dynamic nature, and suggests that certain regions or time periods may face significantly poorer connectivity, while others maintain strong performance. B. Graph Definition The graph construction models the relationships and transitions between gNBs based on the collected connectivity data, forming the foundation for the GCN model. This structure captures the temporal dependencies inherent in 5G handover scenarios. We first preprocess the timestamp data to ensure consistent datetime formatting, enabling precise alignment of consecutive records within the same day. This is essential for defining transitions between gNBs. The resulting graph Gconsists of 26 nodes, representing unique gNBs identified by their gNB IDs, and 79 edges, reflecting potential transitions. Each node is characterized by connectivity-related attributes: •Current connectivity: The average of the last three connectivity values for the gNB (or fewer if limited data is available). This smoothing strikes a balance between capturing recent trends and reducing short-term fluctuations, ensuring robustness without delaying responses to abrupt changes. •Previous connectivity (t−1): The connectivity value from the prior timestamp, introducing a temporal dimension for capturing trends. •Target connectivity for t+ 1: The most recent connectivity value, serving as the prediction target. Edges represent detected handovers in consecutive records from the same day. Each edge is weighted by the absolute difference in connectivity values between connected nodes, capturing variations in connectivity quality during transitions. Larger differences emphasize transitions with a greater impact on connectivity. This approach results in a graph that balances stability and responsiveness in handover decisions, accurately representing the dynamic relationships between gNBs. Figure 2 presents an example of the constructed graph, where each node represents a gNB and each edge signifies a potential connectivity transition. Fig. 2. Example of the constructed graph C. Graph Convolutional Networks We employ GCNs with SAGEConv layers to predict optimal gNB connections based on historical connectivity data. SAGEConv layers are effective for learning from graphstructured data by aggregating information from neighboring nodes. 1) Model Architecture: The architecture of our GCN model consists of several key components: •Input Layer: The model takes as input a feature matrix Xassociated with the graph nodes and an edge index E that defines the graph structure. •Graph Convolutional Layers: Two SAGEConv layers are used, each with a distinct aggregation strategy (e.g., mean or max). The first layer projects input features to a hidden dimension defined by the hidden_channels parameter, followed by a ReLU activation. The second layer refines node representations and outputs the predicted connectivity values. •Output Layer: The model’s output is a single layer that produces a connectivity for each node after passing through a sigmoid activation function, which scales the outputs between 0 and 1. The forward pass of the model proceeds as follows: H(1) =ReLU(SAGEConv(X,E)) (2) H(2) =SAGEConv(H(1),E)(3) connectivity =sigmoid(H(2))(4) ReLU introduces non-linearity in the hidden layer, while sigmoid constrains outputs to the [0,1] range, aligning with the connectivity metric. This architecture leverages node features and graph structure to predict optimal connections within the 5G network. 2) Training Procedure: The training process for our GCN model involves the following steps: •Data Splitting: The dataset is split sequentially into 70% training, 15% validation, and 15% test sets, preserving temporal order to maintain continuity in training and evaluation. •Optimization: We use the Adam and AdamW optimizers during hyperparameter tuning to find the best configuration for learning rate, weight decay and dropout rates. The loss function for connectivity prediction is based on Smooth L1 Loss, with a beta parameter of 0.2 to provide robustness against outliers. •Hyperparameter Tuning: A grid search is conducted over hyperparameters, including hidden dimensions, learning rate, dropout rate, aggregation methods and optimizers. The model with the lowest validation loss and minimal training-validation loss difference is selected to mitigate overfitting. •Training and Validation: During each epoch, the model computes connectivity predictions on the training set, and parameters are updated using backpropagation. Validation loss is monitored after each epoch to evaluate performance. The best model is selected based on validation performance. •Evaluation: The best model is tested on the unseen data, and test loss is computed to assess generalization. Validation and test loss trends are analyzed to evaluate the model stability and performance. D. Handover Decision Strategy with Connectivity Threshold Our handover decision model addresses the ”ping-pong” effect, where frequent handovers between neighboring gNBs degrade the network performance, particularly in high-mobility or dense network scenarios. Traditional signal strength-based methods lack robustness in 5G environments that require consideration of additional metrics like latency and bandwidth. To mitigate this, we implemented a threshold-based strategy leveraging predicted connectivity values at the next time step (t+ 1). Rather than reacting to current connectivity values, the model forecasts t+ 1 connectivity for the current gNB and its neighbors. This enables proactive handover decisions that prioritize stability and optimize connection quality. The handover is triggered when the neighboring gNB has a predicted connectivity at t+1 that is higher than the predicted connectivity of the current gNB, and the difference between these predicted connectivity values exceeds a threshold of 0.05. This threshold value was chosen as a balanced cutoff to avoid unnecessary handovers due to minor fluctuations in connectivity while ensuring responsiveness to meaningful improvements in connectivity. While the threshold offers an effective buffer for stabilizing connections, its precise value may vary depending on network conditions and could be further optimized in future studies. IV. RESULTS AND DISCUSSION This section presents the results of the GCN model training and evaluation, followed by an analysis of the impact of different connectivity weight configurations on the network performance, handover decisions, and latency optimization. A. Hyperparameter Tuning Results To optimize the GCN model, a grid search was conducted over various hyperparameters. Table II summarizes the explored parameter grid and the optimal configuration. TABLE II HYPERPARAMETER TUNING RESULTS Hyperparameter Values Tested / Optimal Value hidden_dim [8, 16, 32, 64] / 16 learning_rate [0.01, 0.001] / 0.01 dropout_rate [0.3, 0.4, 0.5] / 0.3 aggregator [’mean’, ’max’, ’sum’] / ’max’ weight_decay [0.01, 0.005, 0.001] / 0.001 optimizer [’Adam’, ’AdamW’] / ’Adam’ This setup achieved a training loss of 0.0077, a validation loss of 0.0029, and a loss difference of 0.0048, demonstrating effective performance on both seen and unseen data. Figure 3 shows the loss curves for the best model, where the smooth decline and minimal gap between training and validation losses indicate convergence and generalization. Fig. 3. Training and Validation Loss Over Epochs for the Best Model The low validation loss with a modest hidden dimension of 16 highlights the model’s efficiency in capturing connectivity patterns without excessive complexity. The small loss difference (0.0048) reinforces the model’s generalizability, avoiding overfitting. These results validate the chosen hyperparameters, offering a lightweight yet effective architecture suitable for real-time connectivity management in 5G networks. B. Connectivity Prediction and Handover Decision Results To evaluate the model’s handover strategy, we compared the connectivity trajectories of the model’s gNB choices with those from the real gNB sequence. This analysis highlights the model’s ability to optimize connectivity over time. Figure 4 shows the connectivity trajectories over a specific day. The blue line represents the real connectivity values, which exhibit significant fluctuations due to varying network conditions. The red dashed line represents the model’s gNB choices, forming a more stable trajectory by consistently selecting gNBs with higher predicted connectivity. Fig. 4. Connectivity of Real vs Model-Simulated gNB Choices Over a Specific Day Expanding to the entire dataset, Figure 5 confirms the model’s long-term effectiveness, maintaining higher connectivity levels compared to the real trajectory. This demonstrates the robustness of the model in identifying gNBs that offer superior connectivity under diverse temporal conditions. Fig. 5. Connectivity of Real vs Model-Simulated gNB Choices Over the Entire Dataset The model enhances connectivity without increasing handover frequency, thanks to a 0.05 connectivity threshold that prevents unnecessary transitions caused by minor fluctuations. This effectively mitigates the ”ping-pong” effect, ensuring that handovers occur only when substantial connectivity gains are expected, resulting in a stable and efficient handover pattern. By selectively connecting to gNBs with higher predicted connectivity, the model maintains a stable trajectory, prioritizes connectivity quality, and minimizes unnecessary transitions. This approach leads to a more seamless connectivity experience, improving the user satisfaction and network performance. In terms of latency, the model prioritizes gNBs with lower latency as part of the connectivity calculation. Figure 6 compares the average latency of the real trajectory and the model’s gNB choices over the same day as Figure 4, showing that the model consistently achieves lower latency. These results demonstrate that the model effectively balances connectivity and latency, delivering a seamless user TABLE III PERFORMANCE METRICS FOR DIFFERENT WEIGHT CONFIGURATIONS Weights (RTT, RSRP, Bitrate, Loss) Avg. Connectivity Avg. Latency (ms) Handovers (Model) Handovers (Real) RTT=0.4, RSRP=0.2, Bitrate=0.2, Loss=0.2 0.756 69.29 9 12 RTT=0.5, RSRP=0.15, Bitrate=0.2, Loss=0.15 0.782 112.00 17 12 RTT=0.7, RSRP=0.1, Bitrate=0.1, Loss=0.1 0.862 200.53 25 12 RTT=0.25, RSRP=0.25, Bitrate=0.25, Loss=0.25 0.798 85.59 11 12 Fig. 6. Recent Average Latency for Real vs Model-Simulated Trajectories Over the Same Specific Day experience for latency-sensitive mobility applications. By prioritizing gNBs with lower latency and higher connectivity, the model supports critical mobile scenarios, ensuring improved network performance and user satisfaction. C. Impact of Connectivity Weight Configurations To refine the connectivity calculations and assess the impact of different metric weights, we conducted a study over a 3hour period on a specific day. The objective was to evaluate the influence of various weight configurations on predicted connectivity, average latency, and handover frequency. Table III summarizes the results for four configurations, highlighting the trade-offs between stability, latency, and handover efficiency. The configuration wRTT = 0.4balances connectivity stability, low latency, and a handover frequency close to the real trajectory. While its average connectivity is slightly lower than in higher RTT-weighted configurations, it significantly reduces latency and ensures a stable handover process, making it suitable for latency-sensitive applications. In contrast, higher RTT weights (wRTT = 0.7) achieve the highest connectivity scores, but introduce excessive handovers, increasing network instability. Equal weighting (wRTT = 0.25) provides balanced performance, but does not effectively prioritize latency. This analysis underscores the importance of carefully tuning weights in the connectivity formula to align with the specific performance requirements of 5G applications. V. CONCLUSION AND FUTURE WORK This study proposed a GCN-based approach to optimize the handover management in 5G networks by predicting connectivity at the next time step (t+ 1), and incorporating a weighted connectivity score that prioritizes RTT for latency-sensitive applications. The model improves handover decisions by dynamically selecting gNBs with higher predicted connectivity, reducing unnecessary transitions and mitigating the ”ping-pong” effect. The results demonstrate connectivity stability, with the model consistently identifying gNBs that provide superior connection quality compared to the observed real trajectory. By prioritizing gNBs with lower latency, the approach supports applications requiring low-latency connections, while ensuring that handovers occur only when substantial connectivity gains are evident. This threshold-based strategy increases network efficiency, reduces the resource usage, and contributes to improved user experience in highmobility scenarios. Future work will refine the connectivity scoring mechanism by integrating adaptive metrics and optimizing thresholds in real time to dynamically adjust handover criteria based on network conditions. Testing the GCN model in live 5G environments and integrating it with advanced orchestration frameworks like SDN and NFV will provide valuable insights, improve real-time analytics, and boost automation capabilities. These efforts aim to enhance the robustness, adaptability, and practical utility of GCN-based handover strategies, contributing to more efficient and resilient next-generation telecommunication systems. REFERENCES [1] Z. Wu, S. Pan, et al., “A Comprehensive Survey on Graph Neural Networks,” IEEE Trans. Neural Netw. Learning Syst., vol. 32, no. 1, pp. 4–24, Jan. 2021 [2] P. Tam, I. Song, et al., “Graph Neural Networks for Intelligent Modelling in Network Management and Orchestration: A Survey on Communications,” Electronics, vol. 11, no. 20, p. 3371, Oct. 2022 [3] J. Chen, Y. Yang, et al., g, “Multitask Offloading Strategy Optimization Based on Directed Acyclic Graphs for Edge Computing,” IEEE Internet Things J., vol. 9, no. 12, pp. 9367–9378, Jun. 2022 [4] G. Dong et al., “Graph Neural Networks in IoT: A Survey,” Mar. 31, 2022, arXiv: arXiv:2203.15935 [5] T. N. Kipf and M. Wellin, “Semi-Supervised Classification with Graph Convolutional Networks,” Feb. 22, 2017, arXiv: arXiv:1609.02907. [6] S. A. Khan, I. Shayea, et al., “Handover management over dual connectivity in 5G technology with future ultra-dense mobile heterogeneous networks: A review,” Engineering Science and Technology, an International Journal, vol. 35, p. 101172, Nov. 2022 [7] L. Leng, J. Li, et al., “Graph convolutional network-based reinforcement learning for tasks offloading in multi-access edge computing,” Multimed Tools Appl, vol. 80, no. 19, pp. 29163–29175, Aug. 2021 [8] J. Zhang et al., “Fine-Grained Service Offloading in B5G/6G Collaborative Edge Computing Based on Graph Neural Networks,” in ICC 2022 - IEEE International Conference on Communications, Seoul, Korea, Republic of: IEEE, May 2022, pp. 5226–5231 [9] P. Rito et al., ”Aveiro Tech City Living Lab: A Communication, Sensing, and Computing Platform for City Environments,” in IEEE Internet of Things Journal, vol. 10, no. 15, pp. 13489-13510, 1 Aug.1, 2023 [10] Dias, Duarte, et al., (2023). A Dataset of ITS-G5 and Cellular Vehicular Connectivity in Urban Environment. Data in Brief. 52. 109846. 10.1016/j.dib.2023.109846.